[LITE][OPENCL] Fix OpenCL kernel of exp, tanh; Fix layout pass for opencl backend. (#3212)

* Fix OpenCL kernel of exp, tanh,

* Fix layout pass for opencl backend,

* Add how to debug cl kernel. test=develop.

docs/demo_guides/opencl.md

@@ -65,9 +65,11 @@ rm ./lite/api/paddle_use_ops.h
   --arm_os=android \
   --arm_abi=armv8 \
   --arm_lang=gcc \
-  build_test_arm_opencl
+  build_opencl
 ```
 
+注：如果要调试cl kernel，假设已经完成上述脚本编译(已生成cmake文件)。调试只需要修改`./lite/backends/opencl/cl_kernel/`下对应的kernel文件，保存后在项目根目录执行`python ./lite/tools/cmake_tools/gen_opencl_code.py ./lite/backends/opencl/cl_kernel ./lite/backends/opencl/opencl_kernels_source.cc`，该命令会自动将修改后，再切到build目录下执行`make publish_inference`或者你要编译的单测的可执行文件名，cl kernel文件的内容会随着编译自动打包到产物包如 .so 中或者对应单测可执行文件中。
+
 ### 编译产物说明
 
 编译产物位于`build.lite.android.armv8.gcc.opencl`下的`inference_lite_lib.android.armv8.opencl`文件夹内，这里仅罗列关键产物：

lite/core/mir/static_kernel_pick_pass.h

@@ -145,11 +145,12 @@ class StaticKernelPickPass : public mir::StmtPass {
     }
 
     VLOG(4) << "[score(final)]:" << final_score;
-    VLOG(4) << "-------- pick summary --------";
-    VLOG(4) << " ===> winner_place():" << PrecisionToStr(winner_place.precision)
+    VLOG(2) << "-------- pick summary for " << instruct.op_type()
+            << " --------";
+    VLOG(2) << " ===> winner_place():" << PrecisionToStr(winner_place.precision)
             << " " << DataLayoutToStr(winner_place.layout) << " "
             << TargetToStr(winner_place.target);
-    VLOG(4) << " ===> kernel.place():"
+    VLOG(2) << " ===> kernel.place():"
             << PrecisionToStr(kernel.place().precision) << " "
             << DataLayoutToStr(kernel.place().layout) << " "
             << TargetToStr(kernel.place().target);

lite/core/mir/type_layout_cast_pass.cc

@@ -41,8 +41,9 @@ void TypeLayoutTransformPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
     VLOG(4) << "!node->IsStmt():" << !node->IsStmt();
     if (!node->IsStmt() || node->AsStmt().op_type() == "while") continue;
     auto inlinks = node->inlinks;
-    VLOG(4) << "node->AsStmt().desc:" << node->AsStmt().desc
-            << " inlinks.size():" << inlinks.size();
+    VLOG(4) << "============== node->AsStmt().op_type():"
+            << node->AsStmt().op_type() << " inlinks.size():" << inlinks.size()
+            << " ================";
     for (auto* in : inlinks) {
       ComplementInputs(graph.get(), node, in);
     }
@@ -68,13 +69,25 @@ void TypeLayoutTransformPass::ComplementInputs(SSAGraph* graph,
   CHECK(inst.op_info()->GetInputArgname(in_arg_name, &inst_in_tensor_name));
   auto decl_arg_type =
       inst.picked_kernel().GetInputDeclType(inst_in_tensor_name);
+
   CHECK(in->AsArg().type);
-  VLOG(5) << "\n inst_in_tensor_name:" << inst_in_tensor_name
+  VLOG(3) << "\n inst_in_tensor_name:" << inst_in_tensor_name
           << "\n in->AsArg().name:" << in->AsArg().name
           << "\n *in->AsArg().type:" << *in->AsArg().type
           << "\n *decl_arg_type:" << *decl_arg_type
           << "\n inst.op()->DebugString():" << inst.op()->DebugString();
 
+  // TODO(ysh329): conflict if tensor with kARM target but kImageDefault(OpenCL
+  // layout).
+  // not a good judge, but don't find the source of this issue from
+  // static_pick_kernel_pass
+  // to this pass.
+  auto* in_arg_type = const_cast<Type*>(in->AsArg().type);
+  if (in_arg_type->target() == TARGET(kARM) &&
+      in_arg_type->layout() == DATALAYOUT(kImageDefault)) {
+    return;
+  }
+
   if (!DataLayoutCompatible(*in->AsArg().type, *decl_arg_type)) {
     VLOG(4) << "found Layout unmatched tensor: " << in->AsArg().name
             << " for kernel " << inst.op()->DebugString() << " "

lite/kernels/opencl/activation_image_compute.cc

@@ -177,7 +177,7 @@ REGISTER_LITE_KERNEL(
 
 // exp
 REGISTER_LITE_KERNEL(
-    exp_act,
+    exp,
     kOpenCL,
     kFP16,
     kImageDefault,
@@ -195,7 +195,7 @@ REGISTER_LITE_KERNEL(
 
 // tanh
 REGISTER_LITE_KERNEL(
-    tanh_act,
+    tanh,
     kOpenCL,
     kFP16,
     kImageDefault,

lite/kernels/opencl/activation_image_compute_test.cc

@@ -109,13 +109,13 @@ TEST(act_image2d_fp16, compute) {
                     func_name = "sigmoid";
                     break;
                   case 6:  // tanh
-                    func_name = "tanh_act";
+                    func_name = "tanh";
                     break;
                   case 7:  // tanh
                     func_name = "swish";
                     break;
                   case 8:  // tanh
-                    func_name = "exp_act";
+                    func_name = "exp";
                     break;
                 }
                 LOG(INFO) << "func_name: " << func_name;
@@ -307,7 +307,7 @@ USE_LITE_KERNEL(layout, kOpenCL, kAny, kImageDefault, NCHW_to_ImageDefault);
 USE_LITE_KERNEL(layout, kOpenCL, kAny, kNCHW, ImageDefault_to_NCHW);
 
 // exp
-USE_LITE_KERNEL(exp_act, kOpenCL, kFP16, kImageDefault, ImageDefault);
+USE_LITE_KERNEL(exp, kOpenCL, kFP16, kImageDefault, ImageDefault);
 
 // swish
 USE_LITE_KERNEL(swish, kOpenCL, kFP16, kImageDefault, ImageDefault);
@@ -316,7 +316,7 @@ USE_LITE_KERNEL(swish, kOpenCL, kFP16, kImageDefault, ImageDefault);
 USE_LITE_KERNEL(leaky_relu, kOpenCL, kFP16, kImageDefault, ImageDefault);
 
 // tanh act
-USE_LITE_KERNEL(tanh_act, kOpenCL, kFP16, kImageDefault, ImageDefault);
+USE_LITE_KERNEL(tanh, kOpenCL, kFP16, kImageDefault, ImageDefault);
 
 // relu image2d fp16
 USE_LITE_KERNEL(relu, kOpenCL, kFP16, kImageDefault, ImageDefault);